Twitch is facing a wave of criticism after users discovered that their streams, chats and other channel content can be used to train Amazon’s generative AI models unless they manually opt out.
Twitch Support announced the new control on Wednesday, describing it as a setting that allows creators to opt out of having their channel content used to train generative AI models across Amazon.
Using online content to train AI models is now widespread across the industry, but Twitch’s decision to include creators by default became the more controversial part of the rollout.
The backlash intensified during a livestream discussing the change, when Twitch Chief Product Officer Mike Minton addressed calls for an opt-in system by saying: “If this was opt-in, nobody would opt in.”
Streams, clips and chats are covered
Twitch’s FAQ says content eligible for generative AI training can include livestreams, chats, VODs, clips, highlights, as well as text and images on a creator’s channel, while opting out prevents that material from being used in future training of Amazon-developed models designed to generate or synthesize text, audio, images or video.
However, the setting does not block every use of AI on Twitch, as the company said features including AutoMod, captions, recommendations and AI-supported tools for streamer growth or monetization may continue to process content because they are treated separately from generative model training.
Twitch’s policy also extends to chat messages, meaning users who post in another creator’s stream are subject to that creator’s opt-out choice and cannot independently prevent those messages from being used for training.
The controversy puts consent at the center of Twitch’s AI rollout, with much of the criticism focused not on the existence of an opt-out button, but on creators being included in Amazon’s AI training by default unless they actively switch it off.
Twitch is one piece of Amazon’s AI data strategy
Amazon says its foundation models are trained on a mix of licensed, proprietary, open-source and publicly available material.
Its Nova training documentation shows that Amazon-curated datasets can include entertainment material such as video transcripts and game dialogue, alongside online discussions and other conversational content.
Amazon also says its video-generation models learn from videos paired with written descriptions, helping the systems connect written prompts with corresponding visual content.
Twitch content was being considered for Amazon’s AI development well before the new opt-out setting appeared. In 2024, Minton said content from the platform was already being used in experimental AI work at Amazon, though not at production scale.
Amazon has not disclosed which models currently use Twitch content or how much data is involved.
